Object search strategy for service robots with knowledge-based viewpoint selection and hierarchical action decisions

Abstract

Service robots are frequently tasked with searching for target objects relevant to specific operations. However, the dynamic nature of object locations poses significant challenges for precise localization and tracking. To address this, we propose a unified framework for efficient object search and navigation that integrates viewpoint selection, dynamic map construction, and adaptive hierarchical planning. Our method constructs a visual-topological map (VTMap) that fuses prior knowledge, object-room and object-object co-occurrence statistics, and spatial probability distributions modeled via Gaussian Mixture Models (GMM). The robot continuously generates and updates a room-level probability map, enabling systematic selection of optimal viewpoints. This process maximizes the likelihood of target detection while minimizing travel distance through a utility-based strategy. Multimodal sensory observations are represented as graph nodes, with navigation actions encoded as edges, supporting accurate localization and action planning. To complement global planning, we introduce a hierarchical search strategy that unifies long-term exploration objectives with adaptive local exploration informed by imitation learning. The agent dynamically adjusts its search direction by integrating prior experiences with real-time sensory cues. Local exploration is formulated as a partially observable Markov decision process (POMDP), guided by spatial memory and semantic targets. Furthermore, action cost modeling and an auxiliary inflection point prediction task refine the local exploration process, enabling the system to flexibly transition between global and local search strategies. Collectively, these components facilitate robust and efficient object-oriented navigation in complex and dynamic environments.